Equi-Euler GraphNet: An Equivariant, Temporal-Dynamics Informed Graph Neural Network for Dual Force and Trajectory Prediction in Multi-Body Systems

Vinay Sharma, Rémi Tanguy Oddon, Pietro Tesini, Jens Ravesloot, Cees Taal, Olga Fink · 2025

Read Equi-Euler GraphNet: An Equivariant, Temporal-Dynamics Informed Graph Neural Network for Dual Force and Trajectory Prediction in Multi-Body Systems by Vinay Sharma, Rémi Tanguy Oddon, Pietro Tesini, Jens Ravesloot, Cees Taal, Olga Fink (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-13768.txt.

Introduction

Sharma and five co-authors built this prediction network in twenty twenty-five for machines with many moving parts. Digital twins need fast models that forecast both motion and internal forces. Forces matter because they hint at wear and faults before motion degrades. Their network, Equi-Euler GraphNet, treats parts as nodes and contacts as edges.

Two design biases do the heavy lifting. Edge messages act like physical forces and respect rotations and shifts of the whole system. Node updates mimic Euler time stepping, letting distant effects arrive over time. The test rig was a cylindrical roller bearing, modeled in fine detail. The network beat rival graph models and ran about two hundred times faster than the full solver. Rollouts stayed stable over thousands of steps.

Catalog record

CREATORVinay Sharma, Rémi Tanguy Oddon, Pietro Tesini, Jens Ravesloot, Cees Taal, Olga Fink
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE100 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-13768

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Equi-Euler GraphNet: An Equivariant, Temporal-Dynamics Informed Graph Neural Network for Dual Force and Trajectory Prediction in Multi-Body Systems Vinay Sharma, Rémi Tanguy Oddon, Pietro Tesini, Jens Ravesloot, Cees Taal, Olga Fink Abstract Accurate real-time modeling of multi-body dynamical systems is essential for enabling digital twin applications across industries. While many data-driven approaches aim to learn system dynamics, jointly predicting internal loads and system trajectories remains a key challenge. This dual prediction is particularly important for fault detection and predictive maintenance in continuously operating systems. Internal loads–such as contact forces–serve as early indicators of faults, reflecting changes due to wear, misalignment, or material degradation before these affect system motion. These forces also serve as inputs to physics-based degradation models (e.g., crack growth or wear propagation), enabling damage prediction and remaining useful life estimation, with the system state updated through predicted trajectories for continuous monitoring. In this research, we propose Equi-Euler GraphNet, a physics-informed graph neural network (GNN) framework that simultaneously predicts internal forces and global trajectories in multi-body dynamical systems. In this mesh-free framework, nodes represent discrete components of the multi-body system and edges encode their interactions. Equi-Euler GraphNet introduces two novel inductive biases: (1) an equivariant message-passing scheme interpreting edge messages as interaction forces that remain consistent under Euclidean transformations; and (2) a temporal-aware iterative node update mechanism based on Euler integration that updates node states to capture the influence of distant interactions over time. Tailored for cylindrical roller bearings, the framework introduces a structural bias that decouples the dynamics of the inner and outer rings from the constrained dynamics of the rolling elements. Trained on high-fidelity multiphysics simulations, Equi-Euler GraphNet extrapolates well beyond the training distribution, accurately predicting internal loads and system trajectories under extrapolated rotational speeds, loads, and bearing configurations. It outperforms state-of-the-art GNN baselines focused solely on trajectory prediction, delivering stable rollouts over thousands of time steps with minimal error accumulation. Remarkably, it achieves up to a 200x speedup over conventional multiphysics solvers while maintaining comparable accuracy. These results position Equi-Euler GraphNet as an efficient reduced-order model for digital twins, design, and maintenance tasks. Journal: Mechanical Systems and Signal Processing PACS 0000 , 1111 MSC 0000 , 1111 † † highlights: Novel physics-informed graph neural network framework. Dual task of force and trajectory prediction in modeling of bearing dynamics. Extrapolation to out-of-distribution operating conditions. 1 Introduction Multi-body dynamical systems are foundational to a wide range of engineered applications, including gearboxes, bearing assemblies, pumps, and compressors, where mechanical components interact to deliver the intended system functionality. Accurately understanding and modeling these interactions is crucial not only for achieving reliable performance and meeting design objectives, but also for enabling effective health monitoring, predicting remaining useful life (RUL), and minimizing downtime due to unexpected failures [ 1 ] . Traditional condition monitoring strategies typically rely on external observations, such as accelerometers mounted on housings, and infer system health by analyzing global vibration or acoustic signals [ 2 ] . Although these condition monitoring technologies have demonstrated success in many industrial applications, they often only detect faults without estimating their severity or the fault location, nor do they capture the onset of damage early on. In multi-body systems, faults typically initiate at localized contact interfaces—such as gear meshing points [ 3 , 4 , 5 ] or rolling-element–raceway interfaces in bearings [ 6 , 7 , 8 ] . By the time external vibration signals reveal such internal damage, the fault has often advanced to a more severe stage, leaving only a narrow window for preventive action [ 9 ] . Identifying and localizing faults at an earlier stage would enable more cost-effective refurbishment or partial reconditioning—particularly in large-scale systems like wind turbines—significantly reducing downtime and extending operational life. In principle, internal loads at contact interfaces provide more direct and informative indicators of damage initiation and progression [ 7 , 10 ] . Physically grounded damage models – such as Paris’ law for crack propagation [ 11 ] or wear models for surface degradation [ 12 ] – typically take internal load histories as key inputs, enabling accurate and long-range predictions of damage evolution. Accurately tracking these loads thus holds the potential for earlier detection of fault initiation and supports more precise remaining useful life prediction, facilitating timely and proactive maintenance scheduling. A prominent example where accurate internal load estimation is particularly critical is the rolling-element bearing, a fundamental component in nearly all rotating machinery across aerospace, automotive, and industrial applications. Bearings represent a particularly challenging modeling scenario due to the complex, time-varying contact mechanics between rolling elements and raceways [ 13 , 14 , 15 ] , which are further influenced by operating conditions such as rotational speed, external loads, and lubrication conditions. However, direct measurement of internal loads in systems such as bearings remains extremely challenging – particularly in fast-rotating, confined, or lubricated environments. To overcome this limitation, researchers have turned to physics-based modeling approaches, most notably lumped-parameter models [ 16 , 17 , 18 ] and finite element analysis (FEA) [ 19 , 20 ] , to infer internal force distributions. These models integrate available information – such as component geometry, inertial characteristics, and operating conditions – into the governing equations of motion and contact mechanics, typically formulated as partial differential equations. Examples of physics-based condition monitoring include the use of dynamic bearing models to infer stiffness from measured vibration amplitudes and thereby estimate internal load distributions [ 6 ] , as well as the integration of Paris’ law with FEA to track crack growth in helicopter blades with vibration data as input [ 21 ] . Although physics-based approaches can provide high-fidelity predictions of internal forces and resulting dynamics, their practical deployment in real-world settings remains limited. Lumped-parameter models often require extensive calibration to accurately reflect system dynamics and are highly sensitive to input uncertainties; even minor inaccuracies in estimated mass, stiffness, or damping can result in significant prediction errors [ 22 ] . FEA-based methods, while more comprehensive in modeling contact mechanics and elastohydrodynamic phenomena, are computationally intensive and typically unsuitable for continuous monitoring under varying operating conditions [ 23 ] . Furthermore, as physical systems deviate from the nominal assumptions – whether through geometric tolerances, material inconsistencies, or fluctuating loads – both lumped-parameter and FEA models often require re-parameterization or re-meshing, limiting their adaptability for real-time or online applications. Given the limitations of traditional physics-based models, data-driven surrogate models have emerged as a promising alternative [ 24 , 25 , 26 ] . By learning directly on measured or simulated system data, these models can, in principle, approximate the underlying dynamics without requiring explicit parameter calibration, and they tend to be more tolerant to noisy inputs – a common challenge in industrial environments [ 27 ] . Typically, such models generate trajectory predictions in an autoregressive manner – commonly referred to as rollout – where each predicted state is recursively used as input for the next time step. However, purely data-driven approaches, including recurrent neural networks (RNNs) [ 28 ] , Long Short-Term Memory (LSTM) networks [ 29 ] , convolutional neural network (CNN) [ 30 ] -based simulators, and other deep neural network architectures trained on historical system trajectories, often struggle with error accumulation

Read the complete text